Closed hutuo1213 closed 1 year ago
We include auxiliary action_loss
based on the aux_outputs
of the interaction decoder since it has multi-layer outputs. This is similar to DETR (see https://github.com/facebookresearch/detr) in computing object detection loss.
The mentioned interactive loss (loss_proposal) is computed based on the final outputs of the interaction proposal net, which has no aux_outputs
.
Hi, @zyong812 :
I see the following code, does it mean that only the update of the action loss has been made; There is no update of interactive loss(loss_proposal); If I want to use, the format is similar to the action loss?
yaoyaosanqi.